Extracting low-dimensional psychological representations from\n convolutional neural networks

2020/05/28 by Aditi Jha, Joshua C. Peterson, Jha, Aditi +3
Computer Science · #Computer Vision and Pattern Recognition (cs.CV) #Explainable Artificial Intelligence (XAI) #FOS: Biological sciences #FOS: Computer and information sciences #Neurons and Cognition (q-bio.NC)

paper · pdf · doi:10.48550/arxiv.2005.14363

openalex publication_date 2020/05/28 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

Deep neural networks are increasingly being used in cognitive modeling as a\nmeans of deriving representations for complex stimuli such as images. While the\npredictive power of these networks is high, it is often not clear whether they\nalso offer useful explanations of the task at hand. Convolutional neural\nnetwork representations have been shown to be predictive of human similarity\njudgments for images after appropriate adaptation. However, these\nhigh-dimensional representations are difficult to interpret. Here we present a\nmethod for reducing these representations to a low-dimensional space which is\nstill predictive of similarity judgments. We show that these low-dimensional\nrepresentations also provide insightful explanations of factors underlying\nhuman similarity judgments.\n

Related

• Adapting Deep Network Features to Capture Psychological Representations
• Transforming Neural Network Visual Representations to Predict Human Judgments of Similarity
• Evaluating (and improving) the correspondence between deep neural\n networks and human representations
• Improved prediction of behavioral and neural similarity spaces using pruned DNNs
• Improving neural network representations using human similarity judgments